Download Revised version

ON THE SUMS OF TWO CUBES
BRUCE REZNICK AND JEREMY ROUSE
Abstract. We solve the equation f (x, y)3 + g(x, y)3 = x3 + y 3 for homogeneous
f, g ∈ C(x, y), completing an investigation begun by Viète in 1591. The usual
addition law for elliptic curves and composition give rise to two binary operations
on the set of solutions. We show that a particular subset of the set of solutions is
ring-isomorphic to Z[e2πi/3 ].
1. Introduction
In 1591, François Viète published a revolutionary work on algebra which has been
translated into English [10] as The Analytic Art. Viète’s “Zetetic XVIII” [10, p.145]
is:
Given two cubes, to find numerically two other cubes the sum of which
is equal to the difference between those that are given.
Let the two given cubes be B 3 and D3 , the first the greater, the second
the smaller. Two other cubes are to be found, the sum of which is equal
to B 3 − D3 . Let B − A be the root of the first one that is to be found,
and let B 2 A/D2 −D be the root of the second. Forming the cubes and
comparing them with B 3 − D3 , it will be found that 3D3 B/(B 3 + D3 )
equals A. The root of the first cube to be found, therefore, is [B(B 3 −
2D3 )]/(B 3 + D3 ) and of the second is [D(2B 3 − D3 )]/(B 3 + D3 ). And
the sum of the cubes of these is equal to B 3 − D3 .... [So it is if] B
is 2 and D 1: The cube of the root 6 will equal the individual cubes
of 3, 4 and 5. When, therefore, the cubes of 6x and 3x are given, the
cubes of 4x and 5x will appear and the sum of the latter will be equal
to the difference between the former.
Viète worked at the dawn of algebra, when mathematicians were not yet comfortable with negative numbers; his work can be put into somewhat more modern
terminology by setting B = x and D = −y. Viète’s formula then becomes:
(1.1)
3
3
x +y =
x(x3 + 2y 3 )
x3 − y 3
3
+
y(y 3 + 2x3 )
y 3 − x3
3
.
Date: January 31, 2011.
1991 Mathematics Subject Classification. Primary: 11D25, 11G05; Secondary: 14J27, 14K02.
The second author was supported by NSF grant DMS-0901090.
1
2
BRUCE REZNICK AND JEREMY ROUSE
Equation (1.1) is well-known in number theory; its iteration shows that any sum
of two cubes over Q (except those of the form d3 and 2d3 ) has infinitely many such
representations. See for example [5, §13.7, 21.11]. Continuing Viète’s example,
3
3
(1.2)
189 = 63 + (−3)3 = 43 + 53 = − 1256
+ 1265
= ···
61
61
In this paper, we find all solutions to
(1.3)
f 3 (x, y) + g 3 (x, y) = x3 + y 3 ,
where f (x, y) and g(x, y) are homogeneous rational functions over C. Upon finding
a common denominator for (f, g), the equation in (1.3) becomes
(1.4)
p3 (x, y) + q 3 (x, y) = (x3 + y 3 )r3 (x, y),
where p, q, r ∈ C[x, y] are homogeneous polynomials (forms), f = p/r and g = q/r.
The degree of the solution is defined to be deg(p) = deg(q) = 1 + deg(r).
In projective terms,
(1.5)
(f : g : 1) = (p : q : r).
Our principal definition is the following: let
(1.6)
V = {v = (p : q : r) : where p, q, r ∈ C[x, y] are forms and satisfy (1.4)}.
A solution to (1.6) with r 6= 0 is projectively equivalent to (p/r : q/r : 1) and we
will denote solutions of this type by (p/r, q/r) or (f, g). However, there are three
solutions to (1.4) “at infinity” with r = 0, namely (1 : −1 : 0), (1 : −ω : 0), and
(1 : −ω 2 : 0), where
√
3
1
(1.7)
ω := Exp 2πi
=
−
+
i.
3
2
2
We observe that if π is irreducible and π|p, q in (1.4), then π 3 |(x3 + y 3 )r3 , hence,
π 2 |r3 (at least), so π|r. Similarly, if π|p, r, then π|q and if π|q, r, then π|p. Since r is
a common denominator, no two of {p, q, r} have a common factor.
One would ordinarily say that, if x3 +y 3 = f13 +g13 = f23 +g23 and {f13 , g13 } = {f23 , g23 },
then (f1 , g1 ) and (f2 , g2 ) are the same solution; however as “points” on (1.3), each
solution occurs 18-fold : as (ω j f, ω k g) and (ω k g, ω j f ), where j, k ∈ {0, 1, 2}. We shall
call these elements of V the affiliates of (f, g).
We now list v = (p : q : r) ∈ V (up to affiliation) with degree ≤ 12, with the
convention that subscripts given below to p, q, r will be inherited by f = p/r, g = q/r,
and v. (That these are the only such elements will follow from Theorem 1.1.)
Let
√3 i
(1.8)
ζ := ζ12 = Exp πi
= 2 + 2,
6
√
so that ζ + ζ −1 = 3 and ζ 3 + ζ −3 = 0. We note that there are two solutions of
degree 7, the second of which is the complex conjugate of v7 in the table below.
ON THE SUMS OF TWO CUBES
3
Degree Solution
1
p1 = x
q1 = y
r1 = 1
3
p3 = ζ −1 x3 + ζy 3
3
−1 3
q3 = ζx
√ +ζ y
r3 = 3xy
4
p4 = x(x3 + 2y 3 )
q4 = −y(y 3 + 2x3 )
r4 = x3 − y 3
7
p7 = x (x6 + (1 + 3ω)(x3 y 3 + y 6 ))
q7 = y ((1 + 3ω)(x6 + x3 y 3 ) + y 6 )
r7 = x6 + (1 − 3ω)x3 y 3 + y 6
9
p9 = −x9 + 3x6 y 3 + 6x3 y 6 + y 9
q9 = x9 + 6x6 y 3 + 3x3 y 6 − y 9
r9 = 3xy(x6 + x3 y 3 + y 6 )
12
p12 = −3(x3 − y 3 )3 (x3 + y 3 ) − (1 + 2ω)(x3 (x3 + 2y 3 )3 + y 3 (y 3 + 2x3 )3 )
q12 = −3(x3 − y 3 )3 (x3 + y 3 ) − (1 + 2ω 2 )(x3 (x3 + 2y 3 )3 + y 3 (y 3 + 2x3 )3 )
r12 = 6xy(x3 − y 3 )(2x3 + y 3 )(x3 + 2y 3 )
Observe that
ζ
2+ω
ζ −1
2 + ω2
√ =
√ =
(1.9)
,
3
3
3
3
so that f3 , g3 ∈ Q[ω](x, y); see Theorem 1.2(3) below. By setting x = 6 and y = −3
3
3
in v9 , we obtain the rational solution 189 = 219
+ −51
to (1.2).
38
38
We remark that v4 appears in [3, pp.550-1]. In 1877, Lucas (see [3, p.574]) proved
a formula equivalent to v9 ; namely, if x3 + y 3 = Az 3 , then q93 (x, y) + p39 (x, y) =
Az 3 r93 (x, y). In 1878, Lucas (see [3, p.575]) gave the identity
(1.10)
(−x3 + 3x2 y + 6xy 2 + y 3 )3 + (x3 + 6x2 y + 3xy 2 − y 3 )3 =
33 xy(x + y)(x2 + xy + y 2 )3 ,
which Desboves later observed (see [3, p.575]) is equivalent to Lucas’ previous identity
upon taking (x, y) 7→ (x3 , y 3 ).
Even though v3 is not real, its components become real under any map (x, y) 7→
(αx + βy, ᾱx + β̄y). This map is invertible provided αβ̄ is not real. Taking (α, β) =
(ζ −1 , ζ 1 ) in the expression f33 (x, y) + g33 (x, y) = (x3 + y 3 )r33 (x, y) yields (1.10). Thus,
it can be argued that the first “new” solution to (1.3) in the table is v7 . The previous
interest in (1.3) required solutions over Q. We shall show in Theorem 1.2(7) that
rational solutions only occur for square degree. Since the solution of degree 16 arises
from iterating v4 , our first truly new solution over Q has degree 25.
The set V is invariant under a large number of symmetries, and the examples
given above show that v ∈ V itself may be symmetrical. For example, if v(x, y) ∈ V,
4
BRUCE REZNICK AND JEREMY ROUSE
then v(y, x), v(x, y) (the complex conjugate of v(x, y)), v(x, ωy) and all combinations
thereof are also in V. Further, if v = (f, g), v 0 = (f 0 , g 0 ) ∈ V, then there is a natural
composition, implicit already in the iterations of (1.2). To be specific, we define
w = (h, k) = v ◦ v 0 , by
(1.11)
h(x, y) = f (f 0 (x, y), g 0 (x, y)),
k(x, y) = g(f 0 (x, y), g 0 (x, y))
It follows from
h3 (x, y) + k 3 (x, y) = f 3 (f 0 (x, y), g 0 (x, y)) + g 3 (f 0 (x, y), g 0 (x, y))
(1.12)
= (f 0 (x, y))3 + (g 0 (x, y))3 = x3 + y 3
that v ◦ v 0 ∈ V as well. The connections among v3 , (1.10) and v9 are equivalent to
the equation v3 ◦ v3 = v9 ; it turns out that v3 ◦ v4 = v4 ◦ v3 = v12 . The homogeneous
version of composition (which applies to infinite solutions as well) is given as follows.
If v = (p : q : r) and v 0 = (p0 : q 0 : r0 ), then
(1.13) v ◦ v 0 = (p(p0 (x, y), q 0 (x, y)) : q(p0 (x, y), q 0 (x, y)) : r0 (x, y)r(p0 (x, y), q 0 (x, y))).
Thus, d(v ◦ v 0 ) = d(v)d(v 0 ), unless there are common factors in (1.13). It can be
shown directly that this cannot happen, but it will also follow from our main work.
The first principal result of this paper is the following theorem.
Theorem 1.1. Under the usual addition on elliptic curves (see Section 2 for more
details), V is an abelian group isomorphic to Z+Z+Z/3Z. The generators of infinite
order may be taken to be h1 = (x, y) and h2 = (ωx, ωy); the element of order 3 is
h0 = (1 : −ω : 0), a solution of (1.4) “at infinity”. Further, d(mh1 + nh2 + th0 ) =
m2 − mn + n2 .
Moreover, the subgroup V1 = {mh1 + nh2 } is ring-isomorphic to Z[ω] under the
identification R(mh1 + nh2 ) = m + nω, with addition on curves and composition in
V1 corresponding to addition and multiplication in Z[ω].
This result has myriad consequences on the nature of the solutions (f (x, y), g(x, y));
these are collected as our other main theorem.
Theorem 1.2.
(1) If v ∈ V, then (g(x, y))3 = (f (y, x))3 .
(2) If v, v 0 ∈ V, then v ◦ v 0 and v 0 ◦ v are affiliates.
(3) If v ∈ V, then f, g ∈ Q[ω](x, y).
(4) If d(v) = 3d0 , then some affiliate of v can be written as ṽ ◦ v3 , where d(ṽ) = d0 .
Hence there exist forms P, Q, R so that
p(x, y) = P (x3 , y 3 ),
q(x, y) = Q(x3 , y 3 ),
r(x, y) = xyR(x3 , y 3 )
(5) If d(v) = 3d0 + 1, then, up to a permutation of (p, q), there exist forms P, Q, R
so that
p(x, y) = xP (x3 , y 3 ),
q(x, y) = yQ(x3 , y 3 ),
r(x, y) = R(x3 , y 3 )
ON THE SUMS OF TWO CUBES
5
(6) In no case is d(v) ≡ 2 (mod 3) and no monomial appearing in any p, q, r has
an exponent congruent to 2 mod 3.
(7) The real solutions are precisely those of the form v = mh1 for m ∈ Z and
have d(v) = m2 . The solutions of the form v = (f, f¯) are precisely those of
the form mv3 and have d(v) = 3m2 .
(8) If f (d) denotes the number of solutions with degree d (counting each collection
of affiliates as one solution), then
X e
(1.14)
f (d) =
,
3
e|d
·
where 3 denotesQthe usual Legendre symbol. Thus, the degree of any solution
has the form m2 j pj where the pj are primes ≡ 0, 1 (mod 3).
Here is the organization of the paper. In Section 2, we review the addition law
for points on elliptic curves. This endows V with the structure of an abelian group.
We then analyze a subgroup V0 of V and prove that Theorem 1.2 is true for V0 . We
define a ring isomorphism between a particular subset V1 of V (under addition and
composition c.f. (1.11)) and the ring Z[ω]. In Section 3, we prove that another subset
of V, V∞ , is isomorphic to the endomorphism ring of the elliptic curve (1.4) and use
this to prove that V0 = V and V1 = V∞ . In Section 4, we discuss the implication of
these results for a few related Diophantine equations.
We remark that many (but not all) of these results can also be derived in an entirely
elementary way. We shall present this approach in [6].
We also happily acknowledge helpful conversations with Bruce Berndt and Ken
Ono, and we wish to thank the anonymous referee for helpful comments that improved
the exposition.
2. Point addition
A general reference for this section is [9]. The first part of the presentation has
been heavily influenced by [8], where addition is discussed on the curve X 3 + Y 3 = A.
It is implicit in [8] that A ∈ C, although this is formally unnecessary.
Addition is defined on elliptic curves using a few basic rules. If three points P, Q, R
lie on a line, then P + Q + R = 0. This operation can be shown to be associative
and the set of points on the curve forms an abelian group; see, e.g. [9, p.62]. This is
even true if we look at “curves” whose coordinates are rational functions.
We consider the curve
(2.1)
C : X 3 + Y 3 = A,
where for the moment we will be vague about the underlying space for (X, Y ) and
the nature of A 6= 0. The additive inverse is given by
(2.2)
(X, Y ) + (Y, X) = 0,
6
BRUCE REZNICK AND JEREMY ROUSE
where 0 is the additive identity, to be identified below as a point at infinity on
the curve. To find the explicit value of the sum, we parameterize the line through
two points on C. If (X1 , Y1 ), (X2 , Y2 ) ∈ C, (X1 , Y1 ) 6= (X2 , Y2 ), then the condition
λ(X1 , Y1 ) + (1 − λ)(X2 , Y2 ) ∈ C implies that
(λX1 + (1 − λ)X2 )3 + (λY1 + (1 − λ)Y2 )3 = A
= (λ3 + (1 − λ)3 )A + 3λ(1 − λ)A =⇒
(2.3)
λ(1 − λ) λ(X12 X2 + Y12 Y2 ) + (1 − λ)(X1 X22 + Y1 Y22 ) − A = 0
=⇒ λ = 0, λ = 1 or λ =
A − (X1 X22 + Y1 Y22 )
,
(X12 X2 + Y12 Y2 ) − (X1 X22 + Y1 Y22 )
provided X12 X2 + Y12 Y2 6= X1 X22 + Y1 Y22 . Ignoring λ = 0, 1, and assuming this
condition holds, we have
(2.4)
(X1 , Y1 ) + (X2 , Y2 ) + (W, Z) = 0,
where
A(X1 − X2 ) + Y1 Y2 (X2 Y1 − X1 Y2 )
(X12 X2 + Y12 Y2 ) − (X1 X22 + Y1 Y22 )
A(Y1 − Y2 ) + X1 X2 (X1 Y2 − X2 Y1 )
Z=
(X12 X2 + Y12 Y2 ) − (X1 X22 + Y1 Y22 )
W =
(2.5)
and so
(2.6)
(X1 , Y1 ) + (X2 , Y2 ) = (Z, W ),
again provided that the denominator in (2.5) is not zero.
If X12 X2 + Y12 Y2 = X1 X22 + Y1 Y22 = B, say, then (X1 − X2 )3 + (Y1 − Y2 )3 =
A − 3B + 3B − A = 0, hence Y1 − Y2 = −ω j (X1 − X2 ) for j = 0, 1 or 2. If X1 = X2 ,
then Y1 = Y2 . Otherwise, X1 − X2 6= 0 and
0 = X12 X2 + Y12 Y2 − (X1 X22 + Y1 Y22 ) = X1 X2 (X1 − X2 ) + Y1 Y2 (Y1 − Y2 )
(2.7)
= (X1 − X2 )(X1 X2 − ω j Y1 Y2 ) =⇒ X1 X2 − ω j Y1 (Y1 + ω j (X1 − X2 )) = 0
=⇒ −ω j (Y1 + ω j X1 )(Y1 − ω j X2 ) = 0
If Y1 = −ω j X1 , then A = X13 + Y13 = 0. Otherwise, Y1 = ω j X2 and so Y2 = ω j X1 ;
thus, (X2 , Y2 ) = (ω 2j Y1 , ω j X1 ). To summarize: the three instances in which we
cannot add distinct points according to (2.5) and (2.6) are
(2.8)
(X1 , Y1 ) + (Y1 , X1 ),
(X1 , Y1 ) + (ωY1 , ω 2 X1 ),
(X1 , Y1 ) + (ω 2 Y1 , ωX1 ).
Observe that, if (X2 , Y2 ) = (ω 2j Y1 , ω j X1 ), then the numerators of (W, Z) are, for
j = 0, 1, 2,
(2.9)
A(Y1 − ω j X1 ) + ω 2j X1 Y1 (ω j X12 − ω 2j Y12 )
A(X1 − ω 2j Y1 ) − ω j X1 Y1 (ω j X12 − ω 2j Y12 ),
ON THE SUMS OF TWO CUBES
7
respectively, and are in ratio (1 : −w2j ). In other words, (2.5) fails precisely when
the sum would be one of the points at infinity. Accordingly, we add them to the
definition of C and write
(X1 , Y1 ) + (Y1 , X1 ) + (1 : −1 : 0) = 0,
(2.10)
(X1 , Y1 ) + (ωY1 , ω 2 X1 ) + (1 : −ω 2 : 0) = 0,
(X1 , Y1 ) + (ω 2 Y1 , ωX1 ) + (1 : −ω : 0) = 0.
Note that, for example, (X1 : Y1 : 1), (Y1 : X1 : 1), (1 : −1 : 0) all lie on the projective
line (X1 + Y1 )u3 = u1 + u2 .
In view of (2.2), we see that (1 : −1 : 0) is the additive identity and the point
h0 := (1 : −ω : 0) has order 3. Further, we see that
(2.11)
(ωX1 , ω 2 Y1 ) = (X1 , Y1 ) + h0 ,
(ω 2 X1 , ωY1 ) = (X1 , Y1 ) + 2h0 .
We still need to define addition when (X1 , Y1 ) = (X2 , Y2 ). In this case, we construct
the equivalent to the tangent line to the point at (X1 , Y1 ) to make a double point
and decree that the third intersection point will be −2(X1 , Y1 ). Formal implicit
differentiation says that the “slope” to the curve X 3 + Y 3 = A at (X1 , Y1 ) equals
−X12 /Y12 , so we seek λ 6= 0 so that
3
X12
3
(2.12)
(X1 + λ) + Y1 − λ 2
− A = 0.
Y1
Since X13 + Y13 = A, the left-hand side of (2.12) is divisible by λ2 , so
3X1 Y13
X1 (A + Y13 ) −Y1 (A + X13 )
(2.13)
λ= 3
=⇒ −2(X1 , Y1 ) =
,
.
X1 − Y13
X13 − Y13
X13 − Y13
We now set A = x3 + y 3 ∈ C[x, y] and summarize the foregoing discussion of addition. Addition involving points at infinity is specified by (2.2) and (2.11). Otherwise,
If (X1 , Y1 ) 6= (X2 , Y2 ), then (X1 , Y1 ) + (X2 , Y2 ) = (Z, W ), where
(x3 + y 3 )(Y1 − Y2 ) + X1 X2 (X1 Y2 − X2 Y1 )
,
(X12 X2 + Y12 Y2 ) − (X1 X22 + Y1 Y22 )
(x3 + y 3 )(X1 − X2 ) + Y1 Y2 (X2 Y1 − X1 Y2 )
W =
;
(X12 X2 + Y12 Y2 ) − (X1 X22 + Y1 Y22 )
−Y1 (2X13 + Y13 ) X1 (X13 + 2Y13 )
Otherwise, 2(X1 , Y1 ) =
,
.
X13 − Y13
X13 − Y13
Z=
(2.14)
Note that 2(X1 , Y1 ) = (X1 , Y1 ) ◦ (−v4 ), and that in the extract cited in the Introduction, Viète in [10], in effect, chooses a slope for the line to ensure that the cubic
equation for λ would have a double root at λ = 0, rendering its third root easy to
find. (This tangent line was computed almost 100 years before calculus was invented!)
Silverman [8, p.335] explicitly derived (2.13), with regards to elliptic curves of the
form x3 + y 3 = A with A ∈ C, although he did not make the reference to Viète.
8
BRUCE REZNICK AND JEREMY ROUSE
Finally, since ω 2 = λ + (1 − λ)ω for λ = −ω, we observe that
(X1 , Y1 ) + (X1 , ωY1 ) + (X1 , ω 2 Y1 ) = 0
(X1 , Y1 ) + (ωX1 , Y1 ) + (ω 2 X1 , Y1 ) = 0
(2.15)
(X1 , Y1 ) + (ωX1 , ωY1 ) + (ω 2 X1 , ω 2 Y1 ) = 0.
We now specialize this discussion to V. First, we write the affiliates of (f, g) ∈ V
in arrays to clarify these sums to zero over lines. Write (f, g) = e1 and (ωf, ωg) = e2
for short. Then all affiliates can be expressed in terms of e1 , e2 and h0 :
(f, ω 2 g) = e2 + 2h0
(ωf, ω 2 g) = e1 + h0
(f, ωg) = −e1 − e2 + h0
(ω 2 f, ω 2 g) = −e1 − e2
(ω 2 f, ωg) = e1 + 2h0
(ωf, ωg) = e2
(ωf, g) = −e1 − e2 + 2h0
(f, g) = e1
(ω 2 f, g) = e2 + h0
(2.16)
(g, ω 2 f ) = −e2 + 2h0
(ωg, ω 2 f ) = −e1 + h0
(g, ωf ) = e1 + e2 + h0
(g, f ) = −e1
(ω 2 g, ω 2 f ) = e1 + e2
(ω 2 g, ωf ) = −e1 + 2h0
(ωg, ωf ) = −e2
(ωg, f ) = e1 + e2 + 2h0
(ω 2 g, f ) = −e2 + h0 .
We now recall h0 and identify two special points on V:
(2.17)
h0 = (1 : −ω : 0),
h1 = (x, y),
h2 = (ωx, ωy),
and let
(2.18)
V0 = {mh1 + nh2 + th0 : m, n ∈ Z, t ∈ {0, 1, 2}},
where mh1 + nh2 + th0 is the canonical expression for (f, g) ∈ V0 . (We henceforth
reserve m, n, t to the description above.) We also recall the definition of an important
subset of V0 :
V1 = {mh1 + nh2 : m, n ∈ Z}.
(2.19)
For (0, 0) 6= (m, n) ∈ Z2 , let
(2.20)
T (m, n) = {mh1 + nh2 , mh1 + nh2 + h0 , mh1 + nh2 + 2h0 }
denote the (m, n)-trio.
We now begin to describe the ring-isomorphism between V1 and Z[ω] by analyzing
v ◦ w. Our first results apply to V0 as well.
Lemma 2.1. If v = (f, g) ∈ V, then
(2.21) h1 ◦v = v ◦h1 = v,
h2 ◦v = v ◦h2 = ωv,
h2 ◦h2 = ωh2 = ω 2 h1 = −h1 −h2 .
Proof. The first identity is immediate from the definition of composition. For the
second one, note that f and g are homogeneous of degree 1, hence f (ωx, ωy) =
ωf (x, y) and g(ωx, ωy) = ωg(x, y), so
(2.22)
v ◦ h2 = (f (ωx, ωy), g(ωx, ωy)) = (ωf (x, y), ωg(x, y)) = h2 ◦ v.
The final equation follows from the second and (2.15).
ON THE SUMS OF TWO CUBES
9
We now make a simple, but consequential observation about left-distributivity.
Lemma 2.2. If v, v 0 , w ∈ V and ṽ = v + v 0 , then ṽ ◦ w = v ◦ w + v 0 ◦ w. Thus,
(mv + nv 0 ) ◦ w = mv ◦ w + nv 0 ◦ w.
Proof. Suppose w = (f (x, y), g(x, y)). Composition with w amounts to the formal
substitution (x, y) → (f (x, y), g(x, y)); (1.11) shows that substitution this is preserved
by the varying definitions of addition, establishing the first assertion. The second
assertion follows from the first by induction.
Theorem 2.3. If v = (f, g) = mh1 + nh2 + th0 ∈ V0 , then
(2.23) (ωf, ωg) = −nh1 + (m − n)h2 + th0 ,
(ω 2 f, ω 2 g) = (n − m)h1 − mh2 + th0 .
Proof. Lemmas 2.1 and 2.2 imply that
(2.24) (ωf, ωg) = ωv = v◦h2 = mh1 ◦h2 +nh2 ◦h2 +th0 ◦h2 = mh2 +n(−h1 −h2 )+th0 .
The other equation follows from (2.15).
For x = mh1 + nh2 ∈ V1 , define
(2.25)
R(x) = R(mh1 + nh2 ) = m + nω
Note that if v = mh1 + nh2 ∈ V1 , then Theorem 2.3 implies that
(2.26)
R(ωv) = −n + (m − n)ω = ω(m + nω) = ωR(v).
Using (2.16) and Theorem 2.3, once we know the canonical expression for (f, g), we
know the canonical expressions for all of its affiliates. The canonical expressions for
the solutions listed in the introduction are
v1 = h1 , v3 = h1 + 2h2 , v4 = −2h1 ,
(2.27)
v7 = −2h1 − 3h2 , v9 = −3h1 , v12 = −2h1 − 4h2 .
√
Note that R(v1 ) = 1, R(v3 ) = 1 + 2ω = ω − ω 2 = i 3 and R(v4 ) = −2.
It is clear from (2.16) and Theorem 2.3 that each set of 18 affiliates is a union of
the six trios:
T (m, n), T (−m, −n), T (−n, m − n),
(2.28)
T (n, n − m), T (n − m, −m), T (m − n, m),
and every v ∈ V0 is in a trio with some w ∈ V1 . As long as (m, n) 6= (0, 0), the six
trios in (2.28) are distinct.
The argument of Theorem 2.3 extends to give a closed form for composition in V.
Theorem 2.4. If v = mh1 + nh2 + th0 and v 0 = m0 h1 + n0 h2 + t0 h0 , then
(2.29)
v ◦ v 0 = m(m0 h1 + n0 h2 + t0 h0 ) + n(−n0 h1 + (m0 − n0 )h2 + t0 h0 ) + th0
= (mm0 − nn0 )h1 + (mn0 + m0 n − nn0 )h2 + ((m + n)t0 + t)h0 .
Proof. Since v ◦v 0 = mh1 ◦v 0 +nh2 +th0 ◦v 0 , three applications of Lemma 2.2, keeping
(2.21) in mind, give the result.
10
BRUCE REZNICK AND JEREMY ROUSE
We note a crucial implication of Theorem 2.4 for elements of V1 (t = t0 = 0):
(2.30)
R(v ◦ v 0 ) = (mm0 − nn0 ) + ω(mn0 + m0 n − nn0 ) = R(v)R(w).
For y ∈ Z[ω], let N (y) denote the usual norm. We have
(2.31) Φ(m, n) := N (m + nω) = |m + nω|2 = (m + nω)(m + nω 2 ) = m2 − mn + n2 .
Since N (y) = N (±ω j y), we have
(2.32)
Φ(m, n) = Φ(−m, −n) = Φ(−n, m − n) = Φ(n, n − m)
= Φ(n − m, m) = Φ(m − n, −m).
Further, (2.30) implies that
(2.33)
Φ(mm0 − nn0 , mn0 + m0 n − nn0 ) = Φ(m, n)Φ(m0 , n0 );
of course, (2.32) and (2.33) can also be verified directly. It will follow from Theorem 1.1 that V0 = V and
(2.34)
d(mh1 + nh2 + th0 ) = Φ(m, n) := m2 − mn + n2 .
Corollary 2.5. If v, v 0 ∈ V are given as above, then
(2.35)
v ◦ v 0 − v 0 ◦ v = (((m + n)t0 + t) − ((m0 + n0 )t + t0 ))h0
so v ◦ v 0 and v 0 ◦ v are in the same trio. Furthermore, v ◦ v 0 = v 0 ◦ v if and only if
(m + n − 1)t0 ≡ (m0 + n0 − 1)t (mod 3), in particular, if v, v 0 ∈ V1 .
We also remark that Theorem 2.4 and the to-be-proved formula (2.34) combine to
imply that d(v ◦ v 0 ) = d(v)d(v 0 ), so that no cancellation occurs in the composition.
We note one more corollary to 2.4, which follows from the bi-homogeneity in the
pairs of variables (m, n) and (m0 , n0 ) of (2.29) for elements of V1 ; the corollary is not
generally true in V.
Corollary 2.6. For v, v 0 ∈ V1 and r ∈ Z, (rv)◦v 0 = v◦(rv 0 ) = r(v◦v 0 ). In particular,
taking v 0 = h1 , we have rv = v ◦ (rh1 ).
Theorem 2.7. The map R is a ring isomorphism between V1 (with the operations of
point-addition and composition) and Z[ω]. Furthermore, with the appropriate definitions of multiplication by ω and the usual complex conjugation,
(2.36)
R(ωv) = ωR(v),
R(v) = R(v).
Proof. That V1 is a ring with respect to addition and composition follows from
Lemma 2.2 in one direction (and from Corollary 2.5 in the other). We remark that
Corollary 2.5 implies that the full set V itself is not a ring with composition as “multiplication”, because the right-distributive law fails. In particular, if v, v 0 ∈ V then
by (1.13), h0 = h0 ◦ v = h0 ◦ v 0 = h0 ◦ (v + v 0 ), but h0 6= 2h0 .
Clearly, R is a bijection and R(v + w) = R(v) + R(w). If v = mh1 + nh2 and
w = m0 h1 + n0 h2 , then as we have seen in (2.30), R(v ◦ w) = R(v)R(w). That
R(ωv) = ωR(v) was shown in equation (2.26). For the second statement, we needn’t
ON THE SUMS OF TWO CUBES
11
concern ourselves with points at infinity, and the exact formulas of (2.14) imply that
complex conjugation factors through addition, so that
(2.37)
(X1 , Y1 ) + (X2 , Y2 ) = (Z, W ) =⇒ (X1 , Y1 ) + (X2 , Y2 ) = (Z, W ).
Since h1 = h1 and h2 = (ωx, ωy) = (ω 2 x, ω 2 y) = ωh2 = −h1 − h2 , we see that if
v = mh1 + nh2 , then
v = mh1 + n(−h1 − h2 ) =⇒
(2.38)
R(v) = m + n(−1 − ω) = m + nω 2 = m + nω = R(v).
Corollary 2.8. If v = m0 h1 + n0 h2 ∈ V1 and w = m1 h1 + n1 h2 , then there exists
0 +n0 ω
v 0 ∈ V such that v = v 0 ◦ w if and only if m
∈ Z[ω]; that is, if and only if
m1 +n1 ω
(2.39)
m0 m1 + n0 n1 ≡ m0 n1 ≡ m1 n0
(mod m21 − m1 n1 + n21 )
Proof. A routine calculation shows that
m0 + n0 ω
m0 m1 + n0 n1 − m0 n1 + (m1 n0 − m0 n1 )ω
=
.
(2.40)
m1 + n1 ω
m21 − m1 n1 + n21
Corollary 2.9. If v = mh1 + nh2 ∈ V1 , then
(2.41)
v ◦ v̄ = N (R(v))h1 = (m2 − mn + n2 )h1 .
Proof. It follows from (2.38) that R(v)R(v̄) = N (R(v)).
If particular, since g3 = f¯3 , v̄3 = −v3 and we recover that v3 ◦ v3 = v9 . It follows
from (2.38) that mh1 = mh1 for all m ∈ Z; thus Corollary 2.9 implies that each
v ∈ V1 has a “composition multiple” which is real. Observe that mh1 + nh2 and
nh1 + mh2 are not, in general, affiliates, although Φ(m, n) = Φ(n, m).
Corollary 2.10. If v = mh1 + nh2 , then ωv = nh1 + mh2 .
Proof. This follows immediately from n + mω = ω(m + nω).
Corollary 2.11. If v ∈ V1 , then v and v̄ are affiliates if and only if v is an affiliate
of mv1 or mv3 for some integer m.
Proof. This follows from a somewhat tedious comparison of (2.28) for (m, n) and
(n, m). Equality holds if mn(m − n) = 0 or (m + n)(m − 2n)(2m − n) = 0, which
give multiples of v1 and v3 respectively.
We note that mv1 = mv1 and mv3 = −mv3 . It follows that the number of solutions
of degree d, f (d), is even unless d = m2 or d = 3m2 .
We now turn to proving Theorem 1.2 for points in V0 , assuming Theorem 1.1. We
show that assertions (1) through (6) hold for v, w ∈ V1 ; since the components of
v + th0 differ from v by powers of ω, which do not affect the assertions, the claimed
results will also hold for V0 .
12
BRUCE REZNICK AND JEREMY ROUSE
Proof of Theorem 1.2. We start with (1). Let v = mh1 + nh2 and v 0 = (y, x) = −h1 .
Then v ◦ v 0 = (f (y, x), g(y, x)) by (1.11). On the other hand, R(v ◦ v 0 ) = R(v)R(v 0 ) =
−R(v) = −m − nω = R(−v), hence v ◦ v 0 = −v; that is, (f (y, x), g(y, x)) =
(g(x, y), f (x, y)).
Item (2) is Corollary 2.5.
Item (3) follows from Theorem 1.1 together with the observation (from (2.14)) that
if v and v 0 have coefficients in Q(ω), then so does v + v 0 .
To prove (4), let v = mh1 + nh2 and note that d(v) = (m + n)2 − 3mn ≡ 0
(mod 3) implies that m + n ≡ 0 (mod 3). Applying (2.39) from Corollary 2.8 gives
−m − 2n ≡ −2m ≡ −n (mod 3) as a condition for the existence of v 0 ∈ V1 so that
0
0
v = v 0 ◦ v3 , and these are satisfied when 3 | m + n. If v 0 = ( pr0 , qr0 ), then by (1.13),
p(x, y) = p0 (ζ −1 x3 + ζy 3 , ζx3 + ζ −1 y 3 ),
(2.42)
q(x, y) = q 0 (ζ −1 x3 + ζy 3 , ζx3 + ζ −1 y 3 ),
√
r(x, y) = 3 xy r0 (ζ −1 x3 + ζy 3 , ζx3 + ζ −1 y 3 ),
which verifies the asserted shape. (Compare with the earlier discussion of (1.10).)
Next, we prove (5). Since (m + n)2 ≡ d(v) ≡ 1 (mod 3), for one choice of sign
(say +), we have ±v = mh1 + nh2 , where m + n ≡ 1 (mod 3). (The choice of sign
amounts to a possible permutation of f and g.) Let v = (f, g) = (p/r, q/r). Since
(ωx, y) = −h1 − h2 + 2h0 , Theorem 2.4 now implies that
v ◦ (ωx, y) = (n − m)h1 − mh2 + 2(m + n)h0 = (n − m)h1 − mh2 + 2h0 ,
which by (2.16) and Theorem 2.3 equals (ωf, g). In other words,
p(ωx, y) q(ωx, y)
p(x, y) q(x, y)
(2.43)
,
= ω
,
.
r(ωx, y) r(ωx, y)
r(x, y) r(x, y)
Thus, p(ωx, y)r(x, y) = ωp(x, y)r(ωx, y) and q(ωx, y)r(x, y) = q(x, y)r(ωx, y). Since
p(x, y) and r(x, y) are relatively prime, we have p(x, y)|p(ωx, y); since they have the
same degree, it follows that p(ωx, y) = cp p(x, y) for cp ∈ C. Similarly, q(ωx, y) =
cq q(x, y) and r(ωx, y) = cr r(x, y). Since p, q, r 6= 0, examination at any non-zero
monomial shows that each constant is a power of ω and so all powers of x occuring in p
with non-zero coefficient are congruent modulo 3, and similarly for q and r. Since d ≡
1 (mod 3), the choices are p(x, y) = xP (x3 , y 3 ), yP (x3 , y 3 ) or x2 y 2 P (x3 , y 3 ) for some
polynomial P ; q(x, y) = xQ(x3 , y 3 ), yQ(x3 , y 3 ) or x2 y 2 Q(x3 , y 3 ) for some polynomial
Q; and r(x, y) = R(x3 , y 3 ), xy 2 R(x3 , y 3 ) or x2 yR(x3 , y 3 ) for some polynomial R. Since
p and q are relatively prime, there cannot be a common factor of x or y, hence
(p(x, y), q(x, y)) is either (xP (x3 , y 3 ), yQ(x3 , y 3 )) or (yP (x3 , y 3 ), xQ(x3 , y 3 )). Upon
dividing the components of either side of (2.43), we find that
(2.44)
p(ωx, y)
p(x, y)
=ω
,
q(ωx, y)
q(x, y)
ON THE SUMS OF TWO CUBES
13
hence p(x, y) = xP (x3 , y 3 ) and q(x, y) = yQ(x3 , y 3 ); (2.43) now implies that r(x, y) =
r(ωx, y), so r(x, y) = R(x3 , y 3 ).
Item (6) follows immediately from (4) and (5).
To prove (7), note that (2.11) h0 = (1 : −ω : 0) = (1 : −ω 2 : 0) = 2h0 and so (2.11)
and Theorem 2.7 imply that
(2.45)
mh1 + nh2 + th0 = mh1 + n(−h1 − h2 ) + t(2h0 ).
If v is real, then v = v so that nh1 + 2nh2 − th0 = 0, hence n = t = 0. Note also that
if v = (f, g) = rh1 , then f, g ∈ Q(x, y).
Finally, we turn to (8). Since each solution has 18 affiliates and Theorem 1.1
implies that d(mh1 + nh2 + th0 ) = m2 − mn + n2 , t ∈ {0, 1, 2}, we have
∞
∞
∞
X
X
X
2
2
d
(2.46)
1+6
f (d)z =
z m −mn+n .
d=1
m=−∞ n=−∞
It is fairly well-known that
∞
∞
∞ X
X
X
z 3i+2
z 3i+1
m2 −mn+n2
−
.
(2.47)
z
=1+6
3i+1
3i+2
1
−
z
1
−
z
m=−∞ n=−∞
i=0
The equations (2.46) and (2.47) combine to imply (1.14). The identity (2.47) has a
convoluted history, as described by our colleague Bruce Berndt in, for example, [1,
p.78] and [2, pp.196-199], and by Hirschhorn in [4]. Its arithmetical equivalent is a
special case of an 1840 theorem of Dirichlet. It was found independently by Lorenz
and Ramanujan.
It follows from (1.14) that f (pk ) = k + 1 if p ≡ 1 (mod 3); if p ≡ 2 (mod 3),
then f (pk ) equals 0 or 1, depending on whether k is odd or even. Since f (d) is
multiplicative, f (n) > 0 implies that no prime ≡ 2 (mod 3) can appear to an odd
power in the prime factorization of n. We note also that {f (n)} is unbounded as
n → ∞. Since the taxicab number 1729 = 7 · 13 · 19, we have f (1729) = 8: there are
8 solutions to (1.4) in which p and q have degree 1729.
3. Proof of Theorem 1.1
In this section, we will prove Theorem 1.1. We will consider solutions to
E : p3 + q 3 = (x3 + y 3 )r3
where 0 6= p, q, r ∈ C[x, y] are homogeneous polynomials with deg(p) = deg(q) =
deg(r) − 1.
For such a point P = (a(x, y) : b(x, y) : c(x, y)), the map
φP (p : q : r) = (a(p, q) : b(p, q) : c(p, q)r)
is a morphism from E to itself. There are two basic properties that immediately
follow from the definition. First,
(3.1)
φP (x : y : 1) = (a(x, y) : b(x, y) : c(x, y)) = P.
14
BRUCE REZNICK AND JEREMY ROUSE
Second, for any point P , the map φP permutes the points at infinity: (1 : −1 : 0),
(1 : −ω : 0) and (1 : −ω 2 : 0).
Before we continue, we need a lemma.
Lemma 3.1. The point P = (x : y : 1) ∈ V has infinite order.
Proof. Define the homomorphism φ : E → E 0 by setting x = 2 and y = 1. Thus
E 0 : p3 + q 3 = 9r3 .
We have φ((x : y : 1)) = (2 : 1 : 1). Using standard techniques (e.g. Prop. VII.3.1(b)
or Cor. VIII.7.2 in [9]), one can compute that E 0 (Q), the group of rational points on
E 0 , is isomorphic to Z, and is generated by (2 : 1 : 1). It follows that P has infinite
order, since a homomorphic image of P also has infinite order.
Let V∞ = {P ∈ V : φP (0) = 0} be the subgroup of points P ∈ V so that φP
fixes the chosen point at infinity. Recall that any polynomial map φ : E → E with
φ(0) = 0 is called an isogeny. Theorem III.4.8 of [9] implies that if φ is an isogeny, then
φ(P + Q) = φ(P ) + φ(Q). The set of all isogenies from E to itself is denoted End(E)
and is called the endomorphism ring of E. The two ring operations are addition (in
the group law, so (φ1 + φ2 )(R) = φ1 (R) + φ2 (R)), and function composition. Our
approach to proving Theorem 1.1 will be to define a ring structure on V∞ , and prove
that V∞ ∼
= End(E), and finally show that V∞ = V1 = {mh1 + nh2 : m, n ∈ Z}.
Lemma 3.2. For any two points P, Q ∈ V, we have
φP +Q = φP + φQ .
Proof. From (3.1), we have
φP +Q (x : y : 1) = P + Q
= φP (x : y : 1) + φQ (x : y : 1).
Thus, the point (x : y : 1) is sent to 0 under the map φP +Q − φP − φQ . If S =
φP +Q (0) − φP (0) − φQ (0), then F = φP +Q − φP − φQ − S is a morphism from E to
itself that fixes 0. Thus, F is an isogeny. Any morphism between two curves is either
constant, or each point has finitely many preimages. It follows that ker F is either
finite, or all of E. Since F is an isogeny,
F ([3n](x : y : 1)) = [3n]F ((x : y : 1)) = [3n](−S) = [n]([3](−S)) = 0.
Here, and in the rest of the section, [m](p : q : r) is used instead of m(p : q : r) for
clarity. By Lemma 3.1, (x : y : 1) has infinite order, and hence the kernel of F is
infinite. This implies that F is the zero map, and so
φP +Q (R) − φP (R) − φQ (R) = S.
for any R. Setting R = (x : y : 1) we see that S = 0, and φP +Q = φP + φQ .
ON THE SUMS OF TWO CUBES
15
Recall that h0 = (1 : −ω : 0) ∈ V and 2h0 = (1 : −ω 2 : 0). Clearly
φh0 (R) = h0 ,
φ2h0 (R) = 2h0
for all R ∈ V. It follows that for any point P ∈ V, either P , P − h0 or P − 2h0 ∈ V∞ .
Hence,
V∼
= V∞ × hh0 i.
Lemma 3.3. The subgroup V∞ ⊆ V can be given the structure of a ring by defining
P · Q = φP (Q).
Proof. We know that V∞ is an abelian group. We must show that the multiplication
operator is associative and distributive. By (3.1),
φP (x : y : 1) = P,
we have
P · Q = φP (Q) = φP (φQ (x : y : 1)).
Since φS (x : y : 1) = S for any S ∈ V, it follows that φP ·Q (x : y : 1) = P · Q =
φP (φQ (x : y : 1)). Thus, (x : y : 1) is in the kernel of the isogeny φP ·Q − φP ◦ φQ .
By Lemma 3.1, the kernel is therefore infinite and hence φP ·Q = φP ◦ φQ . The
associativity then follows from the fact that function composition is associative. To
prove the distributive law, we use that φP is an isogeny and hence
P · (Q + R) = φP (Q + R)
= φP (Q) + φP (R)
= (P · Q) + (P · R).
Thus, V∞ naturally has the structure of a ring.
Lemma 3.4. The map τ : V∞ → End(E) given by
τ (P ) = φP
is an isomorphism of rings.
Proof. Lemma 3.2 implies that τ (P + Q) = τ (P ) + τ (Q). In the proof of Lemma 3.3,
we showed that τ (P ·Q) = τ (P )◦τ (Q). Thus, τ is a ring homomorphism. If τ (P ) = 0,
then φP = 0 and so φP ((x : y : 1)) = P = 0. Hence, τ is injective.
Conversely, if φ ∈ End(E), and P = φ(x : y : 1), then φ − φP has (x : y : 1) in
its kernel. Thus, the kernel of φ − φP is infinite and hence φ = φP . It follows that
φ = τ (P ) and so τ is surjective.
A similar argument identifying the Mordell-Weil group of an elliptic surface with
the endomorphism ring was given by Frank de Zeeuw in his master’s thesis [11].
Now, we will prove our main result.
16
BRUCE REZNICK AND JEREMY ROUSE
Proof of Theorem 1.1. In light of the fact that
V∼
= V∞ × hT i,
and that V∞ is isomorphic to End(E) by Lemma 3.4, it suffices to determine End(E).
Theorem VI.6.1(b) of [9] states that if E is an elliptic curve defined over a field of
characteristic zero, then End(E) is isomorphic to either Z or an order in an imaginary
quadratic field. Observe that End(E) contains the map defined by
φ((p : q : r)) = (ωp : ωq : r).
Hereafter we will refer to the map φ as [ω]. This map fixes (1 : −1 : 0), satisfies
[ω]3 = 1, and sends (x : y : 1) to (ωx : ωy : 1). It follows that End(E) ∼
= Z[ω] and
V∞ = V1 = h(x : y : 1), (ωx : ωy : 1)i ∼
= Z × Z.
Now, we will prove that d(mh1 + nh2 + th0 ) = m2 − mn + n2 . It suffices to prove
this with t = 0, since mh1 + nh2 is an affiliate of mh1 + nh2 + th0 .
If P := mh1 + nh2 ∈ V∞ , it is easy to see that the degree of P is the same as the
degree of the map φP : E → E. In this case,
φP ((x : y : 1)) = mh1 + nh2
= m(x : y : 1) + n(ωx : ωy : 1)
= [m + nω](x : y : 1).
Thus, φP = [m + nω]. The ring End(E) is endowed with an involution ˆ· that satisfies
\
λ
+ φ = λ̂ + φ̂
λ[
◦ φ = φ̂ ◦ λ̂
φ ◦ φ̂ = [deg φ]
(see Theorem III.6.2 of [9]). This, together with the fact that deg([m]) = m2 implies
that
ˆ = [ω 2 ].
[ω]
This implies that
[m\
+ nω] = [m + nω 2 ]
and so
[deg([m + nω])] = [m + nω][m + nω 2 ]
= [m2 + (mnω + mnω 2 ) + n2 ]
= [m2 − mn + n2 ].
Since the degree of P equals deg φP = deg[m + nω], we have that the degree of P is
m2 − mn + n2 , as desired.
ON THE SUMS OF TWO CUBES
17
4. Related results and open questions
We conclude with a brief discussion of some related Diophantine equations. It is
classically known that if F (x, y) is a binary cubic form, then after an invertible linear
transformation in (x, y), F (x, y) has one of the following three shapes: x3 , x3 +y 3 , x2 y.
It is natural to wonder whether there are solutions to (1.4) in the other two cases.
Theorem 4.1. The equations
(4.1)
p3 (x, y) + q 3 (x, y) = x3 r3 (x, y),
(4.2)
p3 (x, y) + q 3 (x, y) = x2 y r3 (x, y),
have no non-trivial solutions in forms p, q, r ∈ C[x, y].
Proof. Any solution to (4.1) would be a solution to the Fermat equation X n +Y n = Z n
for n = 3 over C[t], upon setting (x, y) = (1, t). The non-existence of such nonconstant solutions was proved by Liouville in 1879. (See the exposition in [7, pp.263265].)
Assume (4.2) has a solution and rewrite as
(4.3)
x2 y r3 = (p + q)(p + ωq)(p + ω 2 q).
P j
Let F = {p+ω j q : j = 0, 1, 2}. Note that F is linearly dependent:
ω (p+ω j q) = 0,
hence any polynomial that divides two elements of F divides the third, and also
divides p and q. Let (p0 , q0 , r0 ) be a solution of (4.3) in which d = deg r0 is minimal.
If d = 0, then p0 and q0 must be linear and the product of the elements in F is x2 y,
hence x must divide two of them, and so x|p0 , q0 , a contradiction. Now suppose d ≥ 1
and suppose π is an irreducible factor of r0 . If π divides two elements of F, then, as
before, π divides p, q and (p0 : π, q0 /π : r0 /π) is a solution to (4.2) of lower degree. It
follows that if π m is a factor of r0 , then π 3m is concentrated in one member of F. We
may thus write r0 = s0 s1 s2 so that s3j |p0 + ω j q0 . Since the degrees of {p0 + ω j q0 } are
equal, (4.3) implies that the three remaining factors, {x, x, y}, are either dispersed,
one to each p0 + ω j q0 , or combined in a single factor. In the first case, we may again
conclude that x|p0 , q0 , and (4.3) implies that x|r0 , a contradiction. In the second case,
suppose without loss of generality that x2 y|p0 + q0 . Then we have p0 + q0 = x2 y s30 ,
p0 + ωq0 = s31 , p0 + ω 2 q0 = s32 , and the linear dependence on the elements of F implies
that
(4.4)
x2 ys30 = −ωs31 − ω 2 s32 ,
which, after the absorption of constants, is a solution to (4.2). If deg s1 = d, then
deg s2 = d, deg s0 = d − 1 and deg r = deg s0 + deg s1 + deg s2 = 3d − 1 > d,
contradicting its supposed minimality and completing the descent.
We now show that Theorem 1.2(4,5) contains, in effect, the solution to two other
Diophantine equations.
18
BRUCE REZNICK AND JEREMY ROUSE
Theorem 4.2. Any solution in forms a, b, c ∈ C[x, y] to either of the equations
(4.5)
a3 (x, y) + b3 (x, y) = xy(x + y) c3 (x, y),
(4.6)
x a3 (x, y) + y b3 (x, y) = (x + y) c3 (x, y)
can be directly derived from a solution to (1.4).
Proof. If (4.5) holds, then by taking (x, y) 7→ (x3 , y 3 ), we see that
a3 (x3 , y 3 ) + b3 (x3 , y 3 ) = x3 y 3 (x3 + y 3 )c3 (x3 , y 3 ),
(4.7)
hence (a(x3 , y 3 ) : b(x3 , y 3 ) : xy c(x3 , y 3 )) ∈ V, and deg(a(x3 , y 3 )) = 3d0 . In the
language of Theorem 1.2(4), we have (a, b, c) = (P, Q, R); again, compare with (1.10).
Similarly, suppose (4.6) holds; take (x, y) 7→ (x3 , y 3 ) to obtain
x3 a3 (x3 , y 3 ) + y 3 b3 (x3 , y 3 ) = (x3 + y 3 )c3 (x3 , y 3 ).
(4.8)
Thus (xa(x3 , y 3 ) : yb(x3 , y 3 ) : c(x3 y 3 )) ∈ V and deg(xa(x3 , y 3 )) = 3d0 + 1, so that in
the language of Theorem 1.2(5), we have (a, b, c) = (P, Q, R).
The subject of equal sums of two cubes has a very long history. For example, the
Euler-Binet formulas (see e.g. [5, §13.7]) give a complete parameterization to the
equation
X3 + Y 3 = U 3 + V 3
(4.9)
over Q, although an examination of the proof in [5] shows that it also applies to any
field F of characteristic zero, such as C(x, y). The parameterization is:
(4.10)
X = λ(1 − (a − 3b)(a2 + 3b2 ),
Y = λ((a + 3b)(a2 + 3b2 ) − 1),
U = λ((a + 3b) − (a2 + 3b2 )2 ),
V = λ((a2 + 3b2 )2 − (a − 3b)),
where a, b, λ ∈ F . One can easily solve for (a, b, λ) for which X = x, Y = y, U =
f, V = g, although the derivation assumes that f 3 + g 3 = x3 + y 3 , so it is unhelpful
in finding solutions to (1.3). Further, these solutions do not necessarily come from
simple choices of (a, b, λ). For example, in the case of (1.1), a computation shows
that (X, Y, U, V ) = (x, y, f4 , g4 ) arises (uniquely) from
(4.11)
a=
2x2 + 5xy + 2y 2
,
2(x2 + xy + y 2 )
b=−
3xy(x + y)
,
2(x3 − y 3 )
λ=−
(x − y)3
.
9xy
If (g4 , f4 ) is taken instead of (f4 , g4 ), then a is a quotient of two quartics, b is a
quotient of two quintics and λ is a quintic divided by a quartic.
Finally, we look at some more general sums of two cubes. If h, k, F ∈ C(x, y),
h3 + k 3 = F , w = (h, k) and v = (f, g) ∈ V, then there is (at least) a one-sided
composition on all solutions to X 3 + Y 3 = F , given by v ◦ w = (f (h, k), g(h, k)). This
follows from
(4.12)
f 3 (h, k) + g 3 (h, k) = h3 + k 3 = F.
ON THE SUMS OF TWO CUBES
19
For example, with F (x, y) = 2x6 − 2y 6 and γ = 21/3 ,
(4.13)
(x2 + xy − y 2 )3 + (x2 − xy − y 2 )3 = (γx2 )3 + (−γy 2 )3 = 2x6 − 2y 6 .
However, there is clearly no v = (f, g) ∈ V so that x2 + xy − y 2 = f (γx2 , −γy 2 ).
Moreover, there are other solutions to
a3 (x, y) + b3 (x, y) = (2x6 − 2y 6 )c3 (x, y).
(4.14)
For example,
(4.15)
a0 (x, y) = x3 +
√i y 3 ,
3
b0 (x, y) = x3 −
√i y 3 ,
3
c0 (x, y) = x,
(and (a0 (y, x), b0 (y, x), −c0 (y, x))) do not arise from composition of either solution of
(4.13) with V. We look forward to finding the complete structure of the solutions to
(4.14).
References
[1] B. C. Berndt, Number Theory in the Spirit of Ramanujan, Student Mathematical Library 34,
American Mathematical Society, Providence, RI, 2006, MR2246314 (2007f:11001).
[2] B. C. Berndt and R. A. Rankin, Ramanujan; Letters and Commentary, History of Mathematics
9. American Mathematical Society, Providence, RI; London Mathematical Society, London,
1995, MR1353909 (97c:01034).
[3] L. E. Dickson, History of the Theory of Numbers, vol II: Diophantine Analysis, Carnegie Institute, Washington 1920, reprinted by Chelsea, New York, 1971, MR0245500 (39 #6807b).
[4] M. D. Hirschhorn, Three classical results on representations of a number, Seminaire
Lotharingien, B42f, 1998.
[5] G. H. Hardy and E. M. Wright, An introduction to the Theory of Numbers, Sixth edition,
revised by D. R. Heath-Brown and J. H. Silverman, Oxford University Press, Oxford, 2008,
MR2445243 (2009i:11001).
[6] B. Reznick and J. Rouse, Viete’s Bolero, in preparation.
[7] P. Ribenboim, 13 Lectures on Fermat’s Last Theorem, Springer-Verlag, New York-Heidelberg,
1979, MR0551363 (81f:10023).
[8] J. H. Silverman, Taxicabs and Sums of Two Cubes, Amer. Math. Monthly, 100 (1993), 331-340,
MR1209462 (93m:11025).
[9] J. H. Silverman, The arithmetic of elliptic curves, Second edition, Graduate Texts in Mathematics, 106, Springer, Dordrecht, 2009, MR2514094 (2010i:11005).
[10] François Viète, The Analytic Art: Nine studies in Algebra, Geometry and Trigonometry from
the Opus Restitutae Mathematicae Analyseos, seu Algebrâ Novâ, translated by T. Richard
Witmer, The Kent State University Press, Kent, Ohio 1983, Reprinted by Dover Publications,
Inc., Mineola, NY, 2006, MR0731262 (86b:01012).
[11] Frank de Zeeuw, An elliptic surface of rank 15, master’s thesis, University of Groningen, 2006.
Available at http://irs.ub.rug.nl/dbi/47ea4d97b56fa.
Department of Mathematics, University of Illinois at Urbana-Champaign, Urbana,
IL 61801
E-mail address: [email protected]
Department of Mathematics, Wake Forest University, Winston-Salem, NC 27109
E-mail address: [email protected]